Research On Aphid Detection Method Based On Deep Learning And Key Points

Pests have always been a factor affecting the yield and quality of crops.How to identify pests quickly,accurately and timely is a key link of pest control,which has a very important impact on agricultural production.At present,the identification of pests can be divided into two categories.One is to rely on agricultural pest experts with excellent professional level to judge pests by observing their appearance,habits and other characteristics.However,this kind of operation depends on the experience of experts very much,and it is easy to identify mistakes to a certain extent,and there are time-consuming and laborious problems.Another relies on computers to identify pests.In recent years,many researchers at home and abroad have done a lot of research on the identification of pests by computer,which provides technical and theoretical support for the intelligent detection of pests.However,there are many problems in pest image such as different scale,different distribution and influence of light,which lead to low accuracy of traditional pest detection task in complex environment such as field.With the upgrade of hardware and the development of deep learning technology in the field of image processing,the target detection task of pest image is widely used in the field of computer vision.In a series of large-scale and different pest target detection tasks achieved good results.Taking aphid as the research object and deep neural network technology as the theoretical basis,this thesis focuses on using convolutional neural network and image keypoint technology to detect aphid.The main research work is as follows:(1)Aiming at the problem of low detection accuracy caused by lighting,background,volume and other factors,an association algorithm based on key points is proposed.This algorithm can effectively alleviate the adverse effects of complex background and light on the accuracy of aphid detection,and can roughly detect the location of aphid.The core idea is to further filter information by treating aphids as key points,and at the same time make category association for the remaining key points.First,transform the color space of aphid dataset and add Gaussian blur.Then,the key points in the image are found by the key points algorithm.Finally,the key points are classified selectively and the rough positions are given by using the correlation point algorithm.Experimental results show that the proposed method can reduce redundant information in images.(2)Aiming at the problem that the target size is too small to extract aphid characteristics well and the correlation points cannot directly participate in the calculation,candidate boxes generation algorithm based on the correlation points is proposed.The main idea is to generate candidate regions by the approximate regions given by the correlation points.Candidate regions were trained as new data sets because aphid traits were easier to pick up.A large number of experimental results show that the feature extraction ability of the model trained bythis algorithm is greatly improved.In addition,an aphid detection method based on coordinate region mapping was proposed to solve the problem that the model was not effective in original image detection.The main idea is to divide the original image into images of similar size to the training set.Then,the subgraphs are tested respectively.Finally,the detection results are re-mapped to the original graph dataset.A large number of experiments show that the proposed method can improve the accuracy of the general target detection framework in aphid data sets.